AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add more apps to 2_Cookbook

Browse Source 
Change-Id: Iafe462df9726a32f450bd240a2de3eaa73a10057
This commit is contained in:
Sandeep Kumar
2016-10-14 18:00:26 +05:30
committed by Maneesh Gupta
parent a46e251daf
commit 04af19866f
21 changed files with 686 additions and 94 deletions
Download Patch File Download Diff File Expand all files Collapse all files
samples/2_Cookbook/6_dynamic_shared/Makefile
+36
View File
@@ -0,0 +1,36 @@
HIP_PATH?= $(wildcard /opt/rocm/hip)
ifeq (,$(HIP_PATH))
HIP_PATH=../../..
endif
HIPCC=$(HIP_PATH)/bin/hipcc
TARGET=hcc
SOURCES = dynamic_shared.cpp
OBJECTS = $(SOURCES:.cpp=.o)
EXECUTABLE=./dynamic_shared
.PHONY: test
all: $(EXECUTABLE) test
CXXFLAGS =-g
CXX=$(HIPCC)
$(EXECUTABLE): $(OBJECTS)
$(HIPCC) $(OBJECTS) -o $@
test: $(EXECUTABLE)
$(EXECUTABLE)
clean:
rm -f $(EXECUTABLE)
rm -f $(OBJECTS)
rm -f $(HIP_PATH)/src/*.o
samples/2_Cookbook/6_dynamic_shared/Readme.md
+47
View File
@@ -0,0 +1,47 @@
## Using Dynamic shared memory ###
Earlier we learned how to use static shared memory. In this tutorial, we'll explain how to use the dynamic version of shared memory to improve the performance.
## Introduction:
As we mentioned earlier that Memory bottlenecks is the main problem why we are not able to get the highest performance, therefore minimizing the latency for memory access plays prominent role in application optimization. In this tutorial, we'll learn how to use dynamic shared memory.
## Requirement:
For hardware requirement and software installation [Installation](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/INSTALL.md)
## prerequiste knowledge:
Programmers familiar with CUDA, OpenCL will be able to quickly learn and start coding with the HIP API. In case you are not, don't worry. You choose to start with the best one. We'll be explaining everything assuming you are completely new to gpgpu programming.
## Simple Matrix Transpose
We will be using the Simple Matrix Transpose application from the previous tutorial and modify it to learn how to use shared memory.
## Shared Memory
Shared memory is way more faster than that of global and constant memory and accessible to all the threads in the block. For In the same sourcecode, we will use the `HIP_DYNAMIC_SHARED` keyword to declare dynamic shared memory as follows:
` HIP_DYNAMIC_SHARED(float, sharedMem) `
here the first parameter is the data type while the second one is the variable name.
The other important change is:
` hipLaunchKernel(matrixTranspose, `
dim3(WIDTH/THREADS_PER_BLOCK_X, WIDTH/THREADS_PER_BLOCK_Y),
dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y),
sizeof(float)*WIDTH*WIDTH, 0,
gpuTransposeMatrix , gpuMatrix, WIDTH);
here we replaced 4th parameter with amount of additional shared memory to allocate when launching the kernel.
## How to build and run:
Use the make command and execute it using ./exe
Use hipcc to build the application, which is using hcc on AMD and nvcc on nvidia.
## More Info:
- [HIP FAQ](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/docs/markdown/hip_faq.md)
- [HIP Kernel Language](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/docs/markdown/hip_kernel_language.md)
- [HIP Runtime API (Doxygen)](http://gpuopen-professionalcompute-tools.github.io/HIP)
- [HIP Porting Guide](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/docs/markdown/hip_porting_guide.md)
- [HIP Terminology](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/docs/markdown/hip_terms.md) (including Rosetta Stone of GPU computing terms across CUDA/HIP/HC/AMP/OpenL)
- [clang-hipify](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/clang-hipify/README.md)
- [Developer/CONTRIBUTING Info](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/CONTRIBUTING.md)
- [Release Notes](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP/RELEASE.md)
samples/2_Cookbook/6_dynamic_shared/dynamic_shared.cpp
+141
View File
@@ -0,0 +1,141 @@
/*
Copyright (c) 2015-2016 Advanced Micro Devices, Inc. All rights reserved.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/
#include<iostream>
// hip header file
#include "hip/hip_runtime.h"
#define WIDTH 16
#define NUM (WIDTH*WIDTH)
#define THREADS_PER_BLOCK_X 4
#define THREADS_PER_BLOCK_Y 4
#define THREADS_PER_BLOCK_Z 1
// Device (Kernel) function, it must be void
// hipLaunchParm provides the execution configuration
__global__ void matrixTranspose(hipLaunchParm lp,
float *out,
float *in,
const int width)
{
// declare dynamic shared memory
HIP_DYNAMIC_SHARED(float, sharedMem);
int x = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
int y = hipBlockDim_y * hipBlockIdx_y + hipThreadIdx_y;
sharedMem[y * width + x] = in[x * width + y];
__syncthreads();
out[y * width + x] = sharedMem[y * width + x];
}
// CPU implementation of matrix transpose
void matrixTransposeCPUReference(
float * output,
float * input,
const unsigned int width)
{
for(unsigned int j=0; j < width; j++)
{
for(unsigned int i=0; i < width; i++)
{
output[i*width + j] = input[j*width + i];
}
}
}
int main() {
float* Matrix;
float* TransposeMatrix;
float* cpuTransposeMatrix;
float* gpuMatrix;
float* gpuTransposeMatrix;
hipDeviceProp_t devProp;
hipGetDeviceProperties(&devProp, 0);
std::cout << "Device name " << devProp.name << std::endl;
int i;
int errors;
Matrix = (float*)malloc(NUM * sizeof(float));
TransposeMatrix = (float*)malloc(NUM * sizeof(float));
cpuTransposeMatrix = (float*)malloc(NUM * sizeof(float));
// initialize the input data
for (i = 0; i < NUM; i++) {
Matrix[i] = (float)i*10.0f;
}
// allocate the memory on the device side
hipMalloc((void**)&gpuMatrix, NUM * sizeof(float));
hipMalloc((void**)&gpuTransposeMatrix, NUM * sizeof(float));
// Memory transfer from host to device
hipMemcpy(gpuMatrix, Matrix, NUM*sizeof(float), hipMemcpyHostToDevice);
// Lauching kernel from host
hipLaunchKernel(matrixTranspose,
dim3(WIDTH/THREADS_PER_BLOCK_X, WIDTH/THREADS_PER_BLOCK_Y),
dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y),
sizeof(float)*WIDTH*WIDTH, 0,
gpuTransposeMatrix , gpuMatrix, WIDTH);
// Memory transfer from device to host
hipMemcpy(TransposeMatrix, gpuTransposeMatrix, NUM*sizeof(float), hipMemcpyDeviceToHost);
// CPU MatrixTranspose computation
matrixTransposeCPUReference(cpuTransposeMatrix, Matrix, WIDTH);
// verify the results
errors = 0;
double eps = 1.0E-6;
for (i = 0; i < NUM; i++) {
if (std::abs(TransposeMatrix[i] - cpuTransposeMatrix[i]) > eps ) {
printf("%d cpu: %f gpu %f\n",i,cpuTransposeMatrix[i],TransposeMatrix[i]);
errors++;
}
}
if (errors!=0) {
printf("FAILED: %d errors\n",errors);
} else {
printf ("dynamic_shared PASSED!\n");
}
//free the resources on device side
hipFree(gpuMatrix);
hipFree(gpuTransposeMatrix);
//free the resources on host side
free(Matrix);
free(TransposeMatrix);
free(cpuTransposeMatrix);
return errors;
}